diff --git a/src/amd/compiler/aco_insert_waitcnt.cpp b/src/amd/compiler/aco_insert_waitcnt.cpp
index 71d073965f8..6e292a9bb72 100644
--- a/src/amd/compiler/aco_insert_waitcnt.cpp
+++ b/src/amd/compiler/aco_insert_waitcnt.cpp
@@ -1082,6 +1082,12 @@ handle_block(Program* program, Block& block, wait_ctx& ctx)
       }
    }
 
+   /* For last block of a program which has succeed shader part, wait all memory ops done
+    * before go to next shader part.
+    */
+   if (block.kind & block_kind_end_with_regs)
+      force_waitcnt(ctx, queued_imm);
+
    if (!queued_imm.empty())
       emit_waitcnt(ctx, new_instructions, queued_imm);
    if (!queued_delay.empty())
@@ -1153,11 +1159,6 @@ insert_wait_states(Program* program)
          in_ctx[current.index] = ctx;
       }
 
-      if (current.instructions.empty()) {
-         out_ctx[current.index] = std::move(ctx);
-         continue;
-      }
-
       loop_progress = std::max<unsigned>(loop_progress, current.loop_nest_depth);
       done[current.index] = true;
 
diff --git a/src/amd/compiler/aco_lower_to_hw_instr.cpp b/src/amd/compiler/aco_lower_to_hw_instr.cpp
index 480ea19dd58..edadca72bb9 100644
--- a/src/amd/compiler/aco_lower_to_hw_instr.cpp
+++ b/src/amd/compiler/aco_lower_to_hw_instr.cpp
@@ -3065,6 +3065,12 @@ lower_to_hw_instr(Program* program)
 
       Builder bld(program, end_with_regs_block);
       bld.sopp(aco_opcode::s_branch, exit_block->index);
+
+      /* For insert waitcnt pass to add waitcnt in exit block, otherwise waitcnt will be added
+       * after the s_branch which won't be executed.
+       */
+      end_with_regs_block->kind &= ~block_kind_end_with_regs;
+      exit_block->kind |= block_kind_end_with_regs;
    }
 }